llama.cpp: Pull Request Adds Day-0 Support for OpenBMB's MiniCPM-V 4.7 Ahead of Model Release
Summary
- • Open pull request adds MiniCPM-V 4.7 support to llama.cpp, from OpenBMB's tc-mb
- • Author says the model will be open-sourced in a few days and wants day-0 support
- • Adds image-tile canvas layout and a canvas M-RoPE with a constant time position
- • Merge pending a Release CI fix; author reports preliminary numerical verification passed
Details
PR #29416 adds MiniCPM-V 4.7 model support to llama.cpp (ggml-org)
Submitted by tc-mb of OpenBMB, the group behind the MiniCPM-V model line. The PR is titled "[Model] Support MiniCPM-V 4.7".
Author wants day-0 adaptation; model to be open-sourced in a few days
The author said the model weights will be released shortly and that the PR is meant to give llama.cpp users support from release day.
New canvas layout struct (mtmd_canvas_layout) added for image tiles
It describes how image tiles are arranged on a canvas for the model's multimodal input path.
Canvas M-RoPE uses a time component held constant across an image
RoPE sections are [11, 11, 10, 0], with the time component in the 4th position slot. The same time value applies to all positions within one image.
GGUF writer gains add_rope_mrope_time_slot(3) to record the time slot
This stores the position of the time component in the model file metadata so the runtime can apply the right M-RoPE layout.
The v47 model structure is the same as v46
Per the author, the new version reuses the prior architecture, which keeps the required code changes small.
Maintainer ngxson added back time-last 3D rope and grid layout
The additions reuse existing infrastructure in the codebase rather than introducing new mechanisms.
Author reports preliminary numerical verification was OK and suggested merging
This is the author's own early check and is described as preliminary, not a full evaluation.
Merge is pending: a maintainer wants the Release CI fixed first
Another maintainer noted the wait for the Release CI fix, and CISC is expected to do the merge. The PR had not been merged as of the source.
Product Launch = new model support, Strategy = release approach, Tech Info = implementation detail, New Tech = new capability, Context = background, Insight = verification status, Industry Update = project/merge status
What This Means
People who run models locally with llama.cpp could be able to use MiniCPM-V 4.7 on release day, instead of waiting for community ports. The PR is still open and waiting on a CI fix before merge, so timing is not guaranteed. The numerical verification so far is preliminary and from the author.